Abstract
Early diagnosis of lung cancer is a critical task in modern medicine. This study presents a comprehensive analysis of the impact of image preprocessing methods on the efficiency of deep learning for classifying lung Computed Tomography (CT) scans. Using the IQ-OTH/NCCD dataset, we systematically evaluated ten different preprocessing techniques, including contrast enhancement, noise reduction, data augmentation, and resolution change. The results demonstrate that standard data augmentation methods lead to a systematic degradation of classification quality, especially for the minority class (Benign). In contrast, image resolution reduction proved to be an optimal strategy, providing comparable classification quality with a significant reduction in computational costs. Contrast enhancement and noise reduction methods demonstrated stable performance at the level of the original data. The obtained results provide practitioners with evidence-based recommendations for building effective medical image preprocessing pipelines.